Synthetic Data Generation for Disclosure Control
- Introduced by
- Donald Rubin
- Year
- 1993
- Type
- Privacy-preserving data synthesis
- Family
- Privacy-preserving analysis
What does it do?
Synthetic data generation is a statistical disclosure limitation technique introduced by Donald Rubin in 1993, in which values in a confidential dataset are replaced by draws from a fitted posterior predictive distribution rather than released directly.
The resulting artificial records preserve the joint statistical structure of the original data while preventing the identification of real individuals, enabling analysts to work with a publicly releasable dataset that behaves like the original for most inferential purposes.
The intuition behind the method
Imagine a sculptor who, instead of lending you a priceless original statue, hands you a perfect plaster cast made from the same mold. You can measure every dimension, test its weight distribution, and study its proportions without risk to the original. Synthetic data works the same way: the model learns the 'shape' of the real data and generates artificial records that share that shape, so researchers get analytically useful material while the real people behind the records stay hidden.
History
Donald Rubin proposed synthetic data generation in a 1993 discussion paper on statistical disclosure limitation, adapting the multiple-imputation framework he had developed for missing data throughout the 1970s–1980s. Roderick Little elaborated the theory shortly after, distinguishing fully synthetic from partially synthetic approaches. During the 2000s the U.S. Census Bureau piloted synthetic public-use files for several surveys. The rise of generative adversarial networks and variational autoencoders after 2014 broadened the toolkit substantially, making deep-learning synthesizers practical for complex tabular and longitudinal data.
When to use it
Use synthetic data generation when a dataset contains sensitive personal or commercially confidential information that legally or ethically cannot be shared in its original form, yet releasing some version of the data would benefit research or policy analysis. It is well-suited to large, complex microdata files where simple masking techniques would destroy analytical utility. Key assumptions are that the synthesis model captures the relevant joint distribution and that the intended analysis is consistent with what the synthesizer was designed to support. It is a poor fit when the research question depends on individual-level records being authentic, when the synthesis model is badly misspecified, or when adversarial re-identification attacks exploiting subtle distributional artifacts are a serious concern. Alternatives include differential privacy, data enclaves, and query-based access systems.
Limitations and considerations
- Inferential validity depends critically on correct specification of the synthesis model; a misspecified model produces synthetic data with biased distributions that mislead analysts.
- Combining rules require knowledge of whether data are fully or partially synthetic and how many replicates were produced, details that must accompany the release.
- Generating high-fidelity synthetic data for variables with extreme skewness, sparse categories, or complex hierarchical structures remains technically demanding.
- Membership inference and attribute disclosure attacks have shown that synthetic datasets are not unconditionally private; formal privacy guarantees require additional mechanisms such as differential privacy.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Releases a fully self-contained dataset that analysts can use with standard software and workflows without special infrastructure.
- Grounded in a rigorous Bayesian multiple-imputation framework, providing valid variance estimation when combining rules are applied correctly.
- Can reproduce complex multivariate relationships, rare subgroup distributions, and longitudinal structures that simple masking methods distort.
- Scales naturally to high-dimensional microdata and can incorporate auxiliary public data to improve the fidelity of the synthesizer.
Applications
- National statistical agencies releasing public-use microdata files from censuses and household surveys while protecting respondent identities.
- Healthcare systems sharing patient records for epidemiological research without exposing protected health information under HIPAA or GDPR.
- Financial institutions producing synthetic transaction datasets for fraud-detection model development and regulatory stress-testing exercises.
- Social science data archives augmenting restricted-access longitudinal studies with synthetic teaching datasets for classroom instruction.
- Technology companies generating synthetic user-behavior logs for training and benchmarking machine-learning models when real logs are too sensitive to use directly.
How it works
How it works
A generative model — ranging from Bayesian parametric models to sequential regression or deep generative networks — is fitted to the confidential original dataset. The goal is to capture the full joint distribution of all variables, including their correlations and marginal characteristics, so that the synthesizer encodes everything an analyst would need for downstream inference.
Formula
p(X_{\text{synth}} \mid X_{\text{obs}}) = \int p(X_{\text{synth}} \mid \theta)\, p(\theta \mid X_{\text{obs}})\, d\thetaRather than using a single point estimate of the model parameters, the synthesis procedure samples a fresh parameter vector from the posterior distribution given the observed data. This Bayesian step propagates parameter uncertainty into each synthetic replicate, preventing over-confident inferences and aligning the synthesis framework formally with Rubin's multiple-imputation theory.
Formula
\theta^{*} \sim p(\theta \mid X_{\text{obs}})Using the drawn parameter vector, an entirely new dataset of the same dimensions as the original is simulated. This step is repeated M times (typically M = 5–20) to produce M independent synthetic replicates, each reflecting a different draw from the posterior and therefore capturing the full uncertainty attributable to both sampling variability and parameter estimation.
Formula
X_{\text{synth}}^{(m)} \sim p(X \mid \theta^{*}), \quad m = 1, \ldots, MAnalysts run their statistical procedure on each of the M synthetic datasets, obtaining M point estimates and variance estimates. These are combined using Rubin's combining rules: the overall estimate is the average of the M estimates, and the total variance adds the within-replicate variance to a between-replicate component scaled by (M+1)/M, yielding valid frequentist confidence intervals and hypothesis tests.
Formula
\bar{Q} = \frac{1}{M} \sum_{m=1}^{M} \hat{Q}^{(m)}, \quad T = \frac{1}{M}\sum_{m=1}^{M} U^{(m)} + \frac{M+1}{M} B_MCommon pitfalls
- Releasing only a single synthetic replicate (M = 1) invalidates the combining-rules framework and produces overconfident standard errors.
- Fitting the synthesis model on a non-representative subsample and then generating synthetic records at population scale introduces systematic bias that analysts may not detect.
- Treating synthetic data as a perfect substitute for real data in analyses that require individual authenticity — such as record linkage or geographic micro-targeting — leads to substantively wrong conclusions.
- Omitting rare but analytically important subgroups from the synthesis model can cause those groups to be absent or severely distorted in the synthetic output.
Frequently asked questions
How many synthetic replicates (M) do I need?
Rubin's combining rules require at least M = 2 replicates to estimate between-replicate variance. In practice, M = 5 to 20 is standard: smaller M inflates variance estimates and reduces statistical power, while very large M offers diminishing returns. The optimal choice depends on how much of the total variance is attributable to synthesis uncertainty relative to within-replicate sampling variability.
Is synthetic data the same as anonymized data under GDPR?
Regulators generally treat synthetic data cautiously. If the synthesis model memorizes or closely approximates individual records, the output may still be considered personal data. True anonymization under GDPR requires that re-identification be reasonably impossible; synthetic data meets this bar only when provably low-membership-inference risk is demonstrated, often through formal differential privacy guarantees layered on top of the synthesis procedure.
Can I use deep generative models instead of Bayesian regression synthesizers?
Yes. GANs, VAEs, and diffusion models have been applied as synthesizers and can capture highly nonlinear dependencies that parametric models miss. However, they lack the principled posterior-sampling interpretation underpinning Rubin's combining rules, so additional work is needed to quantify synthesis uncertainty and ensure that releasing M replicates from a GAN yields valid inferences rather than M near-identical copies of the same generated distribution.
Related methods
Sources
Rubin, D. B. (1993). Statistical disclosure limitation. Journal of Official Statistics, 9(2), 461–468.
Open publication referenceCopied.